Balancing Dynamic Scheduling Overhead to Maximize SDF Performance

Abstract

With each new mobile communication standard, the range of computational complexity increases. For the fifth-generation (5G), it spans over six orders of magnitude. When translated to a series of SDF graphs to represent the baseband modem computation, one SDF for each TTI, every SDF differs from the other in topology and complexity, matching this very range. The tight deadlines for the processing of data packets demand the SDF actors to be as small as possible to allow a high parallelity. Small SDF actors, however, mean a relatively high scheduling overhead per actor, which is especially painful for dynamic scheduling systems. We assume a high payload-to-overhead ratio (~~ 1.0) under which traditional single-threaded scheduling systems will merely be able to service one computation thread. In our approach, the scheduling overhead is decomposed into different stages, and the effort is parallelized to increase the performance a platform can deliver even with a high payload-to-overhead ratio.